FitMyLLM
NVIDIA· Ampere

NVIDIA CMP 170HX 8 GB

Running LLMs on the CMP 170HX 8 GB — the long read: which models fit at which quantisation, and the settings worth changing. · Or what a budget buys

VRAM
8 GB
Entry-level
Bandwidth
1490
GB/s
Models Q4
214/498
43%
7B Q4 speed
~170
Blazing
Model coverage at Q443% of all
Estimated speed· By model size at Q4

Average speeds at Q4 quantization. Actual performance varies by model architecture and context length.

3B
~397
tok/s
7B
~170
tok/s
14B
—
7.9GB needed
32B
—
18.0GB needed
70B
—
39.4GB needed
Measured rig reports

We rent the machine and time every model on it: decode, VRAM peak, concurrency, watts and cost per million tokens. Yours may already be one of them — and two are free to read in full.

See the reports →
Specifications
VRAM
8 GB
Bandwidth
1490 GB/s
FP16 Compute
50.5 TFLOPS
TDP
250W
Memory
HBM2e
Architecture
Ampere
CUDA Cores
4,480
Tensor Cores
280
PCIe
Gen 1 x4
214
Fast models · >30 tok/s
Real-time chat speed
214
Usable · >10 tok/s
Comfortable for all tasks
214
Total compatible
Fit in VRAM at Q4
Rent it in the cloud

Buying CMP 170HX 8 GB costs $15–$40k and isn’t practical for most teams. Spin one up by the hour instead:

Spin up in ~60s. Pay by the second. Cancel anytime.

Some links are affiliate links — we may earn a small commission at no extra cost to you. This helps keep FitMyLLM free and independent.

Compatible models· 214
S
Falcon-H1R Tiny 90M0.09B
Falcon·256K ctx· chat· reasoning
13244
tok/s · 7% VRAM
›
S
SmolLM2 135M0.135B
Smollm·2K ctx· chat
8830
tok/s · 7% VRAM
›
S
nomic-embed-text-v1.5 100M0.14B
Embedding·8K ctx· chat
8514
tok/s · 7% VRAM
›
S
GPT-2 124M0.14B
Gpt2·1K ctx· chat
8514
tok/s · 7% VRAM
›
S
SmolVLM 256M0.256B
Smollm·8K ctx· chat· vision
4656
tok/s · 8% VRAM
›
S
Gemma 3 270M0.27B
Gemma·32K ctx· chat
4415
tok/s · 8% VRAM
›
S
Snowflake Arctic Embed M v2.00.305B
Embedding·8K ctx· embedding· multilingual
3908
tok/s · 8% VRAM
›
S
bge-large-en-v1.5 335M0.335B
Embedding·1K ctx· chat
3558
tok/s · 9% VRAM
›
S
mxbai-embed-large-v10.335B
Embedding·1K ctx· embedding
3558
tok/s · 9% VRAM
›
S
Snowflake Arctic Embed L0.335B
Embedding·1K ctx· embedding
3558
tok/s · 9% VRAM
›
S
LFM2 350M0.35B
Lfm·125K ctx· chat· tool_use
3406
tok/s · 9% VRAM
›
S
SmolLM2 360M0.36B
Smollm·8K ctx· chat
3311
tok/s · 9% VRAM
›
S
GPT-2 Medium 345M0.38B
Gpt2·1K ctx· chat
3137
tok/s · 9% VRAM
›
S
Qwen 2.5 0.5B0.5B
Qwen·32K ctx· chat
2384
tok/s · 10% VRAM
›
S
SmolVLM 500M0.5B
Smollm·8K ctx· chat· vision
2384
tok/s · 10% VRAM
›
S
Falcon-H1 0.5B0.52B
Falcon·128K ctx· chat
2292
tok/s · 10% VRAM
›
S
BGE-M30.568B
Embedding·8K ctx· embedding
2099
tok/s · 10% VRAM
›
S
Snowflake Arctic Embed L v2.00.568B
Embedding·8K ctx· embedding· multilingual
2099
tok/s · 10% VRAM
›
S
Qwen3-Embedding 0.6B0.6B
Embedding·32K ctx· embedding· multilingual
1987
tok/s · 11% VRAM
›
S
Falcon-H1R Tiny 0.6B0.6B
Falcon·32K ctx· chat· reasoning
1987
tok/s · 11% VRAM
›
S
Falcon Perception 0.6B0.6B
Falcon·4K ctx· vision
1987
tok/s · 11% VRAM
›
S
Qwen 1.5 0.5B0.62B
Qwen·32K ctx· chat
1923
tok/s · 11% VRAM
›
S
LFM2 700M0.74B
Lfm·125K ctx· chat· tool_use
1611
tok/s · 12% VRAM
›
S
Qwen3 0.6B0.75B
Qwen·32K ctx· chat· reasoning
1589
tok/s · 12% VRAM
›
S
GPT-2 Large 774M0.81B
Gpt2·1K ctx· chat
1472
tok/s · 12% VRAM
›
S
Qwen 3.5 0.8B0.87B
Qwen·256K ctx· chat· coding· multilingual
1370
tok/s · 13% VRAM
›
S
InternVL3 1B0.94B
Other·32K ctx· chat· vision
1268
tok/s · 13% VRAM
›
S
MiniCPM5 1B1.08B
Minicpm·128K ctx· chat· reasoning· multilingual
1104
tok/s · 14% VRAM
›
S
K2 Horizon 0.9B1.08B
Other·128K ctx· chat· coding· reasoning
1104
tok/s · 14% VRAM
›
S
TinyLlama 1.1B1.1B
Llama·2K ctx· chat
1084
tok/s · 15% VRAM
›
S
LFM2.5-1.2B-Thinking1.2B
Lfm·122K ctx· chat· reasoning· tool_use
993
tok/s · 15% VRAM
›
S
Llama-3.2-1B1.2B
Llama·4K ctx· chat
993
tok/s · 15% VRAM
›
S
LFM2 1.2B1.2B
Lfm·125K ctx· chat· tool_use· multilingual
993
tok/s · 15% VRAM
›
S
Zamba2 1.2B1.2B
Other·4K ctx· chat
993
tok/s · 15% VRAM
›
S
Ling 3.0 Tiny7.89BMoE
Other·128K ctx· chat· coding· reasoning
917
tok/s · 66% VRAM
›
S
EXAONE-4.0-1.2B1.3B
Exaone·64K ctx· chat
917
tok/s · 16% VRAM
›
S
OPT 1.3B1.3B
Opt·2K ctx· chat
917
tok/s · 16% VRAM
›
S
MiniCPM-V 4.61.3B
Other·256K ctx· chat· vision
917
tok/s · 16% VRAM
›
S
DeepSeek Coder 1.3B1.35B
Deepseek·16K ctx· coding
883
tok/s · 16% VRAM
›
S
Phi-1 1.3B1.42B
Phi·2K ctx· coding
839
tok/s · 17% VRAM
›
Next step

Get personalized recommendations.

See ranked models with benchmark scores, run commands, and precise speed estimates for your CMP 170HX 8 GB.

What this card is worth

CMP 170HX 8 GB holds 214 of the models in our catalogue and is, in practice, a IQ4_XS card — the largest it takes is Falcon3-10B at IQ4_XS.

Tokens/sec per $100
—
8B at Q4_K_M, so cards compare like for like
VRAM per $100
—
what memory costs on this card
The biggest it takes
Falcon3-10B10.3B · IQ4_XS7.1 GB105 tok/sest
Bamba 9B v29.78B · Q4_K_M7.2 GB101 tok/sest
RecurrentGemma 9B9.63B · Q4_K_M6.8 GB102 tok/sest
glm-4-9b9.4B · Q4_K_M6.6 GB105 tok/sest
Yi 1.5 9B9B · Q4_K_M6.5 GB110 tok/sest
Yi Coder 9B9B · Q4_K_M6.5 GB110 tok/sest
K2 Horizon 7B9B · Q4_K_M6.7 GB110 tok/sest
Ministral 3 8B8.92B · Q4_K_M6.6 GB111 tok/sest
Against cards of similar memory
M3 Pro (18GB)12 GB$1,5991.5 tok/s per $100
M1 Pro (16GB)11 GB$9992.9 tok/s per $100
M2 Pro (16GB)11 GB$1,2992.2 tok/s per $100
M4 (16GB)11 GB$4994.0 tok/s per $100
CMP 170HX 8 GB8 GB——

Neighbours in memory rather than in price: memory decides whether a card can do the job at all, so two cards of the same size at different prices is the comparison you are making. Ordered by memory, then bandwidth — not by the value column, which is worked out from bandwidth and price alone and therefore rewards a cheap card whatever its software stack does to that bandwidth in practice. Read it as one input, not as a ranking.

Device under test

NVIDIA CMP 170HX 8 GB — 8 GB VRAM.

CMP 170HX 8 GB spec
Brand
NVIDIA
VRAM
8 GB HBM2e
Bandwidth
1490 GB/s
FP16 Compute
50.5 TFLOPS
FP32 Compute
12.6 TFLOPS
CUDA Cores
4,480
Tensor Cores
280
TDP
250 W
Architecture
Ampere
AI capability
214/ 498 models @ Q4

With 8 GB VRAM and 1490 GB/s bandwidth, this GPU handles models up to 9.63B parameters.

Speed ≈ bandwidth / model_size × efficiency. A 7B model at Q4 runs at ~170 tok/s.

Rent it — live prices
Checking what the clouds are charging…
Top models for CMP 170HX 8 GB
214 fit · showing 20
ModelSizeVRAM Q4tok/sAvg
RecurrentGemma 9B9.63B6.4 GB12435.0
glm-4-9b9.4B6.2 GB12720.5
gemma-2-9b9.2B6.1 GB130171.8
Yi 1.5 9B9B6.0 GB13230.3
Yi Coder 9B9B6.0 GB13235.8
K2 Horizon 7B9B6.0 GB13252.4
Ministral 3 8B8.92B5.9 GB13422.6
Ministral 3 8B Reasoning8.92B5.9 GB13469.0
NVIDIA-Nemotron-Nano-9B-v28.9B5.9 GB13444.2
InternLM3 8B Instruct8.8B5.9 GB13542.8
Granite 4.2 8B8.79B5.9 GB13644.9
Gemma 1 7B8.54B5.7 GB14024.7
CodeGemma 7B8.54B5.7 GB14040.2
LFM2 8B A1B8.3B5.6 GB79524.3
Seed-Coder 8B Instruct8.25B5.5 GB14434.1
Seed-Coder 8B Reasoning8.25B5.5 GB14432.9
DeepSeek R1-0528 Qwen3 8B8.2B5.5 GB14536.3
Qwen3-8B8.2B5.5 GB14586.5
Granite 3.0 8B8.17B5.5 GB14636.4
Granite 3.1 8B8.17B5.5 GB14638.6